NormalBayesClassifier classes size
I am using a NormalBayesClassifier trained on PyramidGFTT features and SIFT descriptors. I have 8 classes, each contain a number of images (eg: 423, 1230, 826, etc). I have trained the classifier on a random segment of 400 images from each class, and then test it on the rest. What I want to ask is, because I have a class that is the smallest (423 images), does this have any influence if I am training the classifier on 80% of images of each class? Can this lead to wrong prediction? I am asking this, because I know that the SVM classifier is dependent on the number of images in each class and I am wondering if NormalBayesClassifier is the same.
Wow you are mixing it up for me. First you talk about naive bayes classifier, which is a binary classifier and then you move on to multi class classification and SVM classifiers. Could you clarify it for me?
Sorry, is it more clear now: "I am asking this, because I know that the SVM classifier is dependent on the number of images in each class and I am wondering if NormalBayesClassifier is the same"?
NormalBayesClassifier is indeed dependent on the amount of training data. However, how are you going to use a binary classifier for a multiclass problem?
NormalBayes is binary? Then why they say that it can have multiple classes: "it's an effective classifier that can handle multiple classes, not just two"?
Ok, I understand: it is based on binary decision trees. But it is said that it can classify multiple classes. How is it dependent on the amount of training data?
What would you suggest me to use instead?
Well let me say I never used it for multiclass classification. t works great for binary classification in my cases however. Multiclass is a difficult topic. I suggest multiclass SVMs.
What do you thing of neural networks? In fact I have of different sizes and aspect ratio. Another thing is that the number of feature per image is not always the same, so there may be some problems in SVM. Does NN neads fixed aspect ratio and size?
I am not a favorite user of neural nets. It is very difficult to actually retrieve what it s doing inside. You need a fixed feature size for every kind of classification and learning method. Else it won't just work. Mostly people take the best set of features and limit that to an amount that can always be retrieved.
How can I sort the features and pick just the most representative ones?